In socket direct mode, it's possible to bind any two (maybe four
in future) PCIe devices with IDs like xxxx:xx:xx.x and
yyyy:yy:yy.y. Bonding member interfaces are unnecessary to have
the same PCIe domain/bus/device ID anymore,

Kernel driver uses "system_image_guid" to identify if devices can
be bound together or not. Sysfs "phys_switch_id" is used to get
"system_image_guid" of each network interface.

OFED 5.4+ is required to support "phys_switch_id".

Signed-off-by: Rongwei Liu <rongw...@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viachesl...@nvidia.com>
---
 doc/guides/nics/mlx5.rst               |  4 +++
 doc/guides/rel_notes/release_21_11.rst |  4 +++
 drivers/net/mlx5/linux/mlx5_os.c       | 43 ++++++++++++++++++++------
 3 files changed, 42 insertions(+), 9 deletions(-)

diff --git a/doc/guides/nics/mlx5.rst b/doc/guides/nics/mlx5.rst
index bae73f42d8..b58236e00a 100644
--- a/doc/guides/nics/mlx5.rst
+++ b/doc/guides/nics/mlx5.rst
@@ -464,6 +464,10 @@ Limitations
   - In order to achieve best insertion rate, application should manage the 
flows per lcore.
   - Better to disable memory reclaim by setting ``reclaim_mem_mode`` to 0 to 
accelerate the flow object allocation and release with cache.
 
+- Bonding under socket direct mode
+
+  - Needs OFED 5.4+.
+
 Statistics
 ----------
 
diff --git a/doc/guides/rel_notes/release_21_11.rst 
b/doc/guides/rel_notes/release_21_11.rst
index dfc2cbdeed..2a6cc765c2 100644
--- a/doc/guides/rel_notes/release_21_11.rst
+++ b/doc/guides/rel_notes/release_21_11.rst
@@ -106,6 +106,10 @@ New Features
   * Added DES-CBC, AES-XCBC-MAC, AES-CMAC and non-HMAC algo support.
   * Added PDCP short MAC-I support.
 
+* **Updated Mellanox mlx5 driver.**
+
+  * Added socket direct mode bonding support.
+
 * **Updated NXP dpaa2_sec crypto PMD.**
 
   * Added PDCP short MAC-I support.
diff --git a/drivers/net/mlx5/linux/mlx5_os.c b/drivers/net/mlx5/linux/mlx5_os.c
index 3746057673..1d57b934fc 100644
--- a/drivers/net/mlx5/linux/mlx5_os.c
+++ b/drivers/net/mlx5/linux/mlx5_os.c
@@ -2008,6 +2008,8 @@ mlx5_device_bond_pci_match(const struct ibv_device 
*ibv_dev,
        FILE *bond_file = NULL, *file;
        int pf = -1;
        int ret;
+       uint8_t cur_guid[32] = {0};
+       uint8_t guid[32] = {0};
 
        /*
         * Try to get master device name. If something goes
@@ -2022,6 +2024,8 @@ mlx5_device_bond_pci_match(const struct ibv_device 
*ibv_dev,
        np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
        if (!np)
                return -1;
+       if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
+               return -1;
        /*
         * The Master device might not be on the predefined
         * port (not on port index 1, it is not garanted),
@@ -2050,6 +2054,7 @@ mlx5_device_bond_pci_match(const struct ibv_device 
*ibv_dev,
                char tmp_str[IF_NAMESIZE + 32];
                struct rte_pci_addr pci_addr;
                struct mlx5_switch_info info;
+               int ret;
 
                /* Process slave interface names in the loop. */
                snprintf(tmp_str, sizeof(tmp_str),
@@ -2080,15 +2085,6 @@ mlx5_device_bond_pci_match(const struct ibv_device 
*ibv_dev,
                                tmp_str);
                        break;
                }
-               /* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
-               if (pci_dev->domain == pci_addr.domain &&
-                   pci_dev->bus == pci_addr.bus &&
-                   pci_dev->devid == pci_addr.devid &&
-                   ((pci_dev->function == 0 &&
-                     pci_dev->function + owner == pci_addr.function) ||
-                    (pci_dev->function == owner &&
-                     pci_addr.function == owner)))
-                       pf = info.port_name;
                /* Get ifindex. */
                snprintf(tmp_str, sizeof(tmp_str),
                         "/sys/class/net/%s/ifindex", ifname);
@@ -2105,6 +2101,30 @@ mlx5_device_bond_pci_match(const struct ibv_device 
*ibv_dev,
                bond_info->ports[info.port_name].pci_addr = pci_addr;
                bond_info->ports[info.port_name].ifindex = ifindex;
                bond_info->n_port++;
+               /*
+                * Under socket direct mode, bonding will use
+                * system_image_guid as identification.
+                * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
+                * All bonding members should have the same guid even if driver
+                * is using PCIe BDF.
+                */
+               ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
+               if (ret < 0)
+                       break;
+               else if (ret > 0) {
+                       if (!memcmp(guid, cur_guid, sizeof(guid)) &&
+                           owner == info.port_name &&
+                           (owner != 0 || (owner == 0 &&
+                           !rte_pci_addr_cmp(pci_dev, &pci_addr))))
+                               pf = info.port_name;
+               } else if (pci_dev->domain == pci_addr.domain &&
+                   pci_dev->bus == pci_addr.bus &&
+                   pci_dev->devid == pci_addr.devid &&
+                   ((pci_dev->function == 0 &&
+                     pci_dev->function + owner == pci_addr.function) ||
+                    (pci_dev->function == owner &&
+                     pci_addr.function == owner)))
+                       pf = info.port_name;
        }
        if (pf >= 0) {
                /* Get bond interface info */
@@ -2117,6 +2137,11 @@ mlx5_device_bond_pci_match(const struct ibv_device 
*ibv_dev,
                        DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
                                ifindex, bond_info->ifindex, bond_info->ifname);
        }
+       if (owner == 0 && pf != 0) {
+               DRV_LOG(INFO, "PCIe instance %04x:%02x:%02x.%x isn't bonding 
owner",
+                               pci_dev->domain, pci_dev->bus, pci_dev->devid,
+                               pci_dev->function);
+       }
        return pf;
 }
 
-- 
2.27.0

Reply via email to